Key points
- &Truth, a new entrant co-founded by Stanford professor Clark Barrett, joins Harmonic, Theorem and Axiom Math in selling AI-driven formal verification built largely on Lean.
- Barrett says the key gap in testing and scanning is that teams never know they have found every issue; proofs give a definite answer.
- &Truth names memory-safety checks of C code, open source and closed, as a leading use case.
- Anthropic and Theorem report that newer models verify with fewer tokens, though the process remains costly.
Why it matters for marketers: As agents write more code, teams need a way to confirm it does what the brief says; verification tools could shift human effort toward writing specs.
Formal verification, which uses mathematical proofs to show software meets stated requirements, is gaining attention as a response to low-quality AI output. Latent Space reports that Harmonic, Theorem, Axiom Math and a new company called &Truth are all building products in this area, with Lean, an open-source proof assistant, at the center of most of them.
Clark Barrett, &Truth's co-founder and co-CEO and a Stanford computer science professor, told Latent Space that common agent checks such as tests, self-critique and AI judges are useful but leave doubt. Teams cannot be sure they caught every problem. Lean, he said, acts as a checker that confirms no mistake was made in a proof. &Truth wants to read code in any language and prove whether it meets precise requirements.
Security is a leading use case. Barrett described scanning the world's C code to confirm memory safety and flag the spots that need fixes. Theorem, which has worked with Anthropic, makes a similar security argument. Their joint post says models are making verification feasible at larger scale. Anthropic reported that a newer Opus model handled a set of 36 small tasks at a 50% success rate while spending roughly two-fifths less on tokens than its predecessor. Verification is still token-heavy, but the cost is trending down.
For marketers and content teams building with coding agents, the practical shift is in where human time goes. Barrett expects people to spend more effort defining specifications and reviewing results, and less on reading code line by line. A separate MarkTechPost headline points to related academic work on verifying GPU kernels, though its details were not available.
Watch whether these vendors turn their research into tools that plug into agent workflows, and whether token costs keep falling. Teams using agents to build sites, automations or data pipelines can start by writing clearer requirements, since those would be the input any verification layer needs.
Sources
This brief is our own summary. Read the full reporting at the source.